Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

Computers & Industrial Engineering 174 (2022) 108825

Contents lists available at ScienceDirect

Computers & Industrial Engineering


journal homepage: www.elsevier.com/locate/caie

Supplier selection and order allocation planning using predictive analytics


and multi-objective programming
Samiul Islam a, Saman Hassanzadeh Amin a, *, Leslie J. Wardley b
a
Department of Mechanical and Industrial Engineering, Toronto Metropolitan University, ON, Canada
b
Shannon School of Business, Cape Breton University, NS, Canada

A R T I C L E I N F O A B S T R A C T

Keywords: Supplier Selection and Order Allocation (SSOA) are two critical strategic decisions in supply chain management.
Deep Learning It is challenging to make these decisions when the demand is unknown. Prediction of demand is a complex
Multilayer Perceptron problem as it highly depends on some parameters such as product cost. In this research, a three-stage framework
SWOT Analysis
is proposed to tackle the hurdles of SSOA planning problem. In the first stage, a hybrid deep learning technique
Mixed-Integer Linear Programming
Supplier Selection and Order Allocation
based on multistep Long-Short Term Memory (LSTM) network is developed to determine the future product
demands. The efficiency of the developed model is evaluated using two standard error measuring techniques.
Then, the results are compared with two other forecasting models to have accurate forecast. One of them is
Seasonal Auto-Regressive Integrated Moving Average (SARIMA), and the other one is a deep learning model
named Multilayer Perceptron (MLP). In the second stage, a fuzzy supplier evaluation model based on Strengths,
Weaknesses, Opportunities, and Threats (SWOT) model is developed to consider qualitative criteria. The third
stage fetches the results from the forecasting model in Stage 1, and the results of the fuzzy model in the second
stage. A unique multi-objective programming model is developed to select the best suppliers and to determine
the allocated orders to them. To derive the efficient solutions, the weighted-sum method is used. The application
of the proposed framework is discussed using a real dataset from the Canadian Juice industries. The results of the
performance comparison among the considered forecasting models show that the developed LSTM model can
lead to less forecasting errors compared to the SARIMA and MLP models.

1. Introduction An efficient SCM incorporates both SS and OA processes (Jia et al.,


2020). This study aims to provide an integrated solution method for
Supplier Selection (SS) process allows organizations to assess, rank, Supplier Selection and Order Allocation (SSOA) problem.
and identify prospective suppliers. The evaluations are performed based In the traditional SC, the suppliers’ selection criteria involve cost and
on some qualitative and quantitative criteria. These criteria are pre­ quality criteria (Hofer et al., 2021). Cost, quality, service level, and lead
determined based on business settings and objectives. The SS criteria time are commonly used SS criteria in traditional SC (Alipour et al.,
most often subsume conflicting criteria (e.g., cost and quality). As a 2021; Orji and Ojadi, 2021; Wang et al., 2021). Moreover, the legisla­
result, finding the suitable trade-off points among multiple conflicting tions regarding the environmental issues force companies to adopt
criteria is challenging (Ahmadi and Amin, 2019; Papen and Amin, 2019; environmental-related criteria, known as green criteria (Mojumder and
Chen et al., 2020). Once the prospective suppliers are determined, two Singh, 2021; Hasani et al., 2021).
parties may interpose a contract by deciding the Order Allocation (OA) In SSOA, order quantities are determined from the selected suppliers
that can satisfy the business demand. Imprecise order amounts may based on some factors such as cost. Generally, an objective function is
cause problems for business. Hence, it is critical to determine the OA defined for each factor which can be minimization or maximization.
accurately (Hasan et al., 2020). The combination of SS and OA processes Thus, the SSOA problem falls under Multi-Objective Optimization
initiates the Supply Chain (SC) among both parties. Business gains and (MOO) problem. To solve the SSOA problem, Linear Programming (LP),
customers’ contentment can be achieved through an efficient Supply Integer Linear Programming (ILP), Mixed-Integer Linear Programming
Chain Management (SCM) system (Mohebalizadehgashti et al., 2020). (MILP), Mixed-Integer Non-Linear Programming (MINLP), and Goal

* Corresponding author.
E-mail addresses: samiul.islam@ryerson.ca (S. Islam), saman.amin@ryerson.ca (S.H. Amin), Leslie_Wardley@cbu.ca (L.J. Wardley).

https://doi.org/10.1016/j.cie.2022.108825
Received 21 March 2022; Received in revised form 5 September 2022; Accepted 16 November 2022
Available online 22 November 2022
0360-8352/© 2022 Elsevier Ltd. All rights reserved.
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

Programming (GP) are some widely adopted models by the researchers 2.1. Supplier selection
(Mohammed et al., 2019; Bektur, 2020; Feng and Gong, 2020; Jia et al.,
2020; Wong, 2020; Esmaeili-Najafabadi et al., 2021; Nasr et al., 2021; Suitable suppliers are selected based on some predefined evaluation
Ventura et al., 2021). As the OA is inextricably tied to the demand, criteria through the SS process. These criteria are determined based on
inaccurate assumptions of demand may cause the SS and OA processes to the business goals and the related legislations. Few literature review
be erroneous (Yousefi, 2021). Prediction of future demand precisely is papers were published in recent years considering different proposed
required for efficient SSOA planning (Islam and Amin, 2020; Liou et al., approaches for SS models (e.g., Tirkolaee et al., 2020; Alipour et al.,
2021). However, to the best of our knowledge, predictive analytical 2021; Wu et al., 2021). Rashidi et al. (2020) published a meta literature
techniques such as Machine Learning (ML) and Deep Learning (DL) have review article regarding the sustainable SS. The authors revealed that
not been implemented by researchers in the SSOA problems to forecast only a few investigations have addressed the social criteria during the SS
uncertain parameters. process, and suppliers’ evaluation in a dynamic environment requires
To address the challenges of the SSOA planning problem, a three- further investigation. Different decision-making techniques for SS indi­
stage methodology is proposed in this study. A hybrid multistep Long- cated that the Analytic Hierarchy Process (AHP) and Linear Program­
Short Term Memory (LSTM) network is designed in the initial stage to ming (LP) are two widely accepted methods to solve the SS problem
forecast the future demands. Afterwards, a fuzzy supplier assessment (Chai and Ngai, 2020).
method based on the Strengths, Weaknesses, Opportunities, and Threats SS criteria can be stated in quantitative or qualitative forms (Liu
(SWOT) model is developed in the second stage. In Stage 3, the best et al., 2019b). Quantitative criteria are expressed using numeric values
suppliers are chosen, and the orders are assigned to them based on both which are determined from the suppliers’ historical data or based on
qualitative and quantitative factors, and by using a multi-objective facts (Islam, Amin, & Wardley, 2021). On the other hand, linguistic
programming model. The weighted-sum method is employed to find expression is used to state the qualitative data that pose uncertainty in
the efficient solutions. The major research contributions of this study are the data (Chen et al., 2020). The uncertainty in the supplier evaluation
summarized as follows: criteria makes the ranking process challenging (Sazvar et al., 2021).
Some quantitative criteria can also be uncertain such as delivery time.
• A new three-stage Supplier Selection & Order Allocation (SSOA) The most common sources of uncertainty from the literature can be
planning framework is designed where demand forecasting is com­ classified as demand, capacity, cost, delivery time, quality, and disrup­
bined with a SSOA planning mathematical model. To the best of our tions. Researchers have used some methods to cope with uncertainty in
knowledge, this is a new integrated framework in the SSOA literature this field. It is evident from the literature that Fuzzy technique is the
that connects demand forecasting and fuzzy SWOT with an SSOA most used method to address uncertainty in this area. Apparently, none
optimization model to tackle uncertainty in quantitative and quali­ of the previous papers used forecasting techniques to predict uncertain
tative criteria, simultaneously. parameters which is one of the goals of this paper.
• In Stage 1, a modified hybrid multistep LSTM network is developed
to get accurate forecasting results. This approach is new where the 2.2. Machine learning for supplier selection
fluctuations components are separated and normalized using Z-score
normalization for better forecasting accuracy. We compare the per­ Machine Learning (ML) algorithms have been used by a few re­
formance of the developed method with two other well-known time searchers for SS. A hybrid Pareto Genetic Algorithm (GA) was developed
series forecasting techniques, including Seasonal Auto-Regressive by Du et al. (2015) to solve a bi-objective problem for the SS process.
Integrated Moving Average (SARIMA), and Multilayer Perceptron Nepal and Yadav (2015) applied Bayesian Network (BN) to quantify risk
(MLP). factors in a supply chain. Then, they used a Decision Tree (DT) algorithm
• In Stage 2, a new fuzzy SWOT model is proposed to deal with both for classifying the suppliers. Sarkis and Dhavale (2015) proposed the
qualitative and quantitative factors under an uncertain environment integration of BN with Monte Carlo–Markov Chain simulation to rank
to evaluate the suppliers. This unique approach enables us to tune the suppliers using three selection criteria from the business operation
the internal and external factors as the parameters in the optimiza­ areas (namely product quality, on-time delivery, and cost variance). The
tion model to analyze and solve the SSOA problem, efficiently. authors also defined three selection criteria for the environmental
• In Stage 3, a new Multi-objective Mixed-Integer Linear Programming concern areas (e.g., energy efficiency, penalties for environmental vio­
(MOMILP) model is developed where demand forecasting along with lations, and pollution control), and three selection criteria from social
the internal and external factors are considered. This integration responsibility (called turnover rate, responsibility to the community,
using multiple sources, products, and periods is new in the MOMILP and philanthropic contributions). In another paper, Hosseini and Barker
literature. Then, the model is solved using the weighted-sum method. (2016) utilized BN to classify and select the suppliers based on tradi­
tional, green, and resilience criteria.
This study is presented using the following structure. Section 2 is Medhi and Mondal (2016) proposed a hybrid model to cluster the
devoted to the literature review. Then, the problem definition and the suppliers by integrating Artificial Neural Network (ANN) with Support
dataset descriptions are discussed in Section 3. Section 4 includes the Vector Machine (SVM). First, ANN was applied to classify the suppliers.
proposed solution framework with a numerical experiment, considering Then, SVM was used to make clusters of those classified suppliers.
a dataset. Section 5 contains the discussions. Lastly, in Section 6, con­ Tavana et al. (2016) integrated AHP method with an ML algorithm
clusions are discussed. called Multi-Layer Perceptron (MLP). AHP was used first to determine
the selection criteria. Then, MLP was applied to rank the suppliers. In the
2. Literature review paper of Jabbarzadeh et al. (2018), the suppliers were clustered based
on their performances using k-mean clustering method. Then, a bi-
To address the Supplier Selection & Order Allocation (SSOA), some objective optimization model was used to select the best suppliers.
methodologies have been proposed, and some models have been Cavalcante et al. (2019) clustered the suppliers based on their risk
developed in the past few years. Compared to the number of published profile using k-Nearest Neighbor (k-NN) method. They selected the
articles on standalone Supplier Selection (SS) models, fewer research suppliers from the lowest risk cluster. Bahadori et al. (2020) proposed an
works have focused on the SSOA problem. The literature review section integrated model using ANN and Fuzzy VIKOR. They classified the
of this paper is divided into three Sub-sections, namely a) Supplier se­ suppliers using ANN based on six criteria and used Fuzzy VIKOR model
lection, b) Machine learning for supplier selection, and c) Supplier se­ to rank them. Recently, SVM was used in the paper of Liou et al. (2021)
lection and order allocation. to determine the selection criteria. Then, SVM was integrated with Fuzzy

2
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

BWM and TOPSIS to rank and select the suppliers. Table 1 includes the economic, and social criteria have been considered. Jia et al. (2020)
ML techniques proposed in the literature for the SS process and the proposed a Goal Programming (GP) approach for solving the SSOA
limitations of those techniques. It is evident that ML techniques have problem. They emphasized the importance of determining the demand
been used for either classification or clustering the suppliers. However, uncertainty for making effective decisions. Ventura et al. (2021) used
forecasting of uncertain parameters using ML techniques has not been two MINLP models to determine the set of suppliers and the optimal
addressed in the SS processes. order quantity per cycle from them. The authors stated two limitations
of their proposed model. The first one is the SS features where only price
and capacity features have been used to select the suppliers. The other
2.3. Supplier selection and order allocation
limitation is the value of demand, which is based on an assumption.
Chance-constrained programing method has been used in some
Instead of addressing just the SS problem, a few researchers have
research to deal with stochastic demand. As an example, Moheb-Aliza­
proposed the integration of SS with the order allocation problem by
deh and Handfield (2018) proposed an integrated chance-constrained
considering it as a single SSOA problem. In a literature review paper,
stochastic model to solve sustainable SSOA problem where lognormal
Schramm et al. (2020) presented that TOPSIS and AHP are two
distribution function is used to determine the demand. Ahmadi and
commonly used approaches, proposed by researchers to solve SSOA
Amin (2019) proposed an integrated chance-constrained stochastic
problem. The review article about quantitative resilience supply chains
model in a closed-loop supply chain where cumulative distribution
(Hosseini et al., 2019) suggests that the pre-determination of suppliers’
function was used to determine the demand. Joint probability distri­
capacity and allocated orders are important to tackle disruption effi­
bution is also used in related research by Jia et al., (2020). However, in a
ciently. Most recently, Naqvi and Amin (2021) published a literature
multiproduct scenario, the seasonality, trend, and demand correlations
review article about SSOA, where deterministic and uncertain optimi­
among products may not be preserved if demand is determined using
zation models have been studied. The authors advised that the uncertain
distribution functions (Lyu et al., 2021). Table 2 comprises the SSOA
parameters of the suppliers’ selection models can be forecasted using
models developed in recent research articles. This table clearly depicts
advanced machine learning techniques as future research. Alegoz and
that the determination of demand is crucial in SSOA problem. Prede­
Yapicioglu (2019) developed a SSOA model by combining fuzzy AHP
termination of demand is hard as demand is an uncertain parameter.
and Goal programming techniques where economic and environmental
Some researchers considered demand as a fixed value, and others
factors were used to evaluate the suppliers. As a future work, the authors
assumed the demand value using different uncertainty handling tech­
suggested adding more supplier evaluation factors such as social and
niques. However, these assumptions may not be accurate which may
technological factors. Esmaeili-Najafabadi et al. (2019) proposed a
lead to inaccurate SSOA results. As the order allocation is directly linked
Mixed-Integer Nonlinear Programming (MINLP) model for optimizing
with the demand, accurate forecasting of demand is crucial which is
SSOA in a centralized supply chain while taking disruption risks into
addressed in this research. It is evident that none of the SSOA literature
account. The authors assumed deterministic values for the SS
has employed ML techniques to determine the uncertain parameters.
parameters.
Zandieh and Aslani (2019) considered economic and environmental
3. Problem statement
factors as the selection criteria. However, they did not consider the ef­
fects of uncertainty on those criteria. Bektur (2020) developed a fuzzy
The challenges for the SSOA processes can be classified into three
optimization model to solve the SSOA problem where environmental,
different contexts. In the first context, one supplier can meet the de­
mands of multiple products. In the second context, multiple suppliers are
Table 1
available to meet the demand of a single product. In this case, the best
A review of supplier selection models using machine learning techniques.
supplier(s) should be selected, and the related orders should be deter­
Authors ML models The ML model Problems mined. The third context contains multiple suppliers for multiple
was used for
products, where the major challenges are selection of suitable supplier
Du et al. Genetic Algorithm Classification Computationally (s) for different products and to determine the allocated orders from
(2015) (GA) expensive. Always may
them. This paper focuses on this case.
not reach the optimal
solution. This study considers the SSOA planning for a Canadian juice
Nepal and Bayesian Classification Class imbalance is a manufacturing company. According to Statistics Canada, the country’s
Yadav Networks (BN) + major issue in DT. juice sector’s revenue in 2021 reached US$3,511 million (Statistics
(2015) Decision Tree Canada, 2021a). In addition, a 4.92 % compound annual growth rate
(DT)
Sarkis and BN + Monte Carlo Classification Model is not tested
(CAGR 2021–2026) is expected in the juice market of Canada (Statista
Dhavale Markov Chain using numerical data. Juice Report, 2021). The company in this study wants to cover 20 % of
(2015) (MCMC) the total citrus juice demand in Ontario for three types of citrus fruit
juices, namely, orange juice, grapefruit juice, and lemon juice. There are
Hosseini and BN Classification Expert judgement is
60 available citrus fruit farms across Ontario (OFVGA Directory, 2021).
Barker required.
(2016) Some of the farms (suppliers) produce only one type of citrus fruit. On
Medhi and Artificial Neural Clustering Large data for ANN is the other hand, a few farms produce all three types of citrus fruits. The
Mondal Network (ANN) + required. production capacities of different suppliers vary based on the area of
(2016) Support Vector cultivable land of each farm. According to the food consumption data of
Machine (SVM)
Tavana et al. Multi-Layer Classification Human vagueness leads
Ontario, the yearly demand for each type of juice changes from year to
(2016) Perceptron (MLP) inaccurate MLP ranking. year (Statistics Canada, 2021b). As the demands of products inherit
Jabbarzadeh k-mean clustering Clustering Outlier data heavily various elements such as seasonality and trend, reliable demand pre­
et al. (2018) affect the model. diction is a key topic in this research.
Cavalcante k-Nearest Clustering Computationally
The studied company is an intermediary juice manufacturing com­
et al. (2019) Neighbor (k-NN) expensive.
Bahadori et al. ANN Classification ANN does not provide pany, between the fruit farms and retail industries. Production data of
(2020) guaranteed our studied company reflects that 1.99 tonnes (thousand kilograms) of
convergence. citrus fruits are required to produce one kiloliter (thousand liters) of raw
Liou et al. SVM Classification Run time is high. juice. However, during citrus juice manufacturing, 50 % of raw juice is
(2021)
mixed with 50 % of water and other substances. Thus, to manufacture

3
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

Table 2
Classification of some SSOA models from the literature.
Authors Supplier Order Problem scenario Uncertainty handling Demand Proposed models
Selection Allocation Multi- Multi- Multi-
sources products periods

Abdel-Baset et al. (2019) Yes (Y) Y TriNs ANP + VIKOR


Alegoz and Yapicioglu Y Y Y Y TrFN + AHP Assumed FGP
(2019)
Alikhani et al. (2019) Y Y IT-2-FS DEA + VIKOR
Esmaeili-Najafabadi et al. Y Y Y Y Assumed MINLP
(2019)
Hosseini et al. (2019) Y Y Y Assumed BOMIP
Kellner and Utz (2019) Y Y Y Fixed MPT + MILP
Liu et al. (2019a) Y Y TFN + AHP TOPSIS + LP
Liu et al. (2019b) Y Y IVITFNs QFD + VIKOR
Lu et al. (2019) Y Y TFN + AHP MCDM
Memari et al. (2019) Y Y IFN TOPSIS
Mohammed et al. (2019) Y Y Y Y Fzzy AHP Assumed Fuzzy TOPSIS + LP
Moheb-Alizadeh and Y Y Y Y Y Assumed MILP + DEA
Handfield (2019)
Zandieh and Aslani (2019) Y Y Y AHP Fixed BWM + Hybrid GA
Bektur (2020) Y Y Y Y Y TFN + AHP Fuzzy LP + TOPSIS
Chen et al. (2020) Y Y TFN + rough fuzzy DMATEL + TOPSIS
numbers
Feng and Gong (2020) Y Y Y Y Assumed LEWM + MOLP
Hasan et al. (2020) Y Y TFN TOPSIS
Hendiani et al. (2020a) Y Y IT-2-FS Likelihood-based
model
Hendiani et al. (2020b) Y Y TFMF Fuzzy BWM
Jia et al. (2020) Y Y Y Y Assumed FGP
Kilic and Yalcin (2020) Y Y IFN Fuzzy TOPSIS + GP
Tirkolaee et al. (2020) Y Y FANP DEMATEL + TOPSIS
Wen et al. (2020) Y Y 2TLWA VIKOR
Wong (2020) Y Y Y Y Assumed GP
Alavi et al. (2021) Y Y TFN Fuzzy BWM
Alipour et al. (2021) Y Y PF set PF-entropy-SWARA-
COPRAS
Esmaeili-Najafabadi et al. Y Y Y Assumed PSO + MINLP
(2021)
Firouzi and Jadidi (2021) Y Y Y Y TrFN Assumed MOMILP
Kaur and Singh (2021) Y Y Y Y Y Fuzzy AHP Assumed DEA + MIP
Mina et al. (2021) Y Y TFN Fuzzy AHP + Fuzzy
TOPSIS
Nasr et al. (2021) Y Y Y Y Y TFN Simulated Fuzzy BWM + MILP
Orji and Ojadi (2021) Y Y TFN + AHP MULTIMOORA
Wang et al. (2021) Y Y TFN MULTIMOORA
Wu et al. (2021) Y Y TFN + IVPFST + GRA + TOPSIS +
EWM DEMATEL
Ventura et al. (2021) Y Y Y Y Assumed MINLP
Proposed model Y Y Y Y Y LSTM þ TrFN Forecasted LSTM þ SWOT þ
MOMILP

Analytic Hierarchy Process (AHP), Analytical Network Process (ANP), Best Worst Method (BWM), Bi-Objective Mixed-Integer Programming (BOMIP), Complex
Proportional Assessment (COPRAS), Data Envelopment Analysis (DEA), Decision-making Trial and Evaluation Laboratory (DEMATEL), Entropy Weighted Method
(EWM), Fuzzy Analytic Network Process (FANP), Fuzzy Goal Programming (FGP), Genetic Algorithm (GA), Goal Programming (GP), Grey Relational degree (GRA),
Interval Type-2 Fuzzy Set (IT-2-FS), Intuitionistic Fuzzy Number (IFN), Interval-valued Intuitionistic Trapezoidal Fuzzy Numbers (IVITFNs), Interval-valued Py­
thagorean Fuzzy Set Theory (IVPFST), Linear Programming (LP), Linguistic Entropy Weight Method (LEWM), Markowitz Portfolio Theory (MPT), Mixed-Integer
Programming (MIP), Mixed-Integer Linear Programming (MILP), Mixed-Integer Non-Linear Programming (MINLP), Multi-Layer Perceptron (MLP), Multi-Objective
Linear Programming (MOLP), Multi-Objective Mixed-Integer Linear Programming (MOMILP),Multi-objective Optimization by Ratio Analysis (MULTIMOORA),
Multi Criteria Decision-Making (MCDM), Particle Swarm Optimization (PSO), Pythagorean Fuzzy (PF), Quality Function Deployment (QFD), Step-wise Weight
Assessment Ratio Analysis (SWARA), Strengths, Weaknesses, Opportunities, and Threats (SWOT), Techniques for Order Preferences by Similarity to Ideal Solution
(TOPSIS), Trapezoidal Fuzzy Numbers (TrFN), Triangular Fuzzy Numbers (TFN), Triangular Neutrosophic Numbers (TriNs), 2-tuple Linguistic Weighted Average
(2TLWA), VIekriterijumsko KOmpromisno Rangiranje (VIKOR).

one kiloliter of citrus juice, approximately-one tonne of citrus fruits is This study utilizes the juice consumption and supply data from the
required. The shelf-life of manufactured juice from our studied company Canadian Government data repository (Statistics Canada, 2021b), which
is five months for orange juice and grapefruit juice, and six months for contains historical juice consumption records since the year 1960. In
lemon juice. The company would like to predict the demands of three addition to the consumption and supply quantities, the dataset contains
types of citrus fruit juices for the next eight quarters of the incoming some attributes including ‘Production’, ‘Beginning stocks’, ‘Imports’,
years. Each quarter includes three months. Depending on the demand ‘Domestic disappearance’, ‘Total disposition’, ‘Manufacturing’, ‘Ex­
prediction of each quarter, the company aims to select suitable supplier ports’, ‘Waste’, and ‘Ending stocks’. The unit values of these attributes
(s) to fulfill one-fifth of that demand and tries to assign proper order are in tonnes. Fig. 1 shows the three different juice consumption data
quantities to the selected fruit supplier(s)/farms. Some criteria such as used in this study. It is observed that there is no steady trend for the juice
cost, quality, capacities of suppliers, suppliers’ distances, delivery per­ consumption. Moreover, the sudden spikes and drops in the figure reveal
formance, and responsiveness are important in this problem. irregular fluctuations in the data.

4
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

Fig. 1. Ontario-based citrus juice consumption by year.

To replace the null values of the studied dataset during the data solve the SSOA planning problem.
cleaning process, the average values of four consecutive quarters are Stage 1 of Fig. 3 contains three forecasting methods used to forecast
used. Two of those four quarters are from the previous two data points, demand. One of them is a novel modified LSTM network. These three
and the rests are the values for the next two quarters. To get a deeper forecasting methods are compared based on their forecasting accuracy.
understanding of the studied data, a correlational plot is used as shown In this study, the Root Mean Square Error (RMSE) and the Mean Abso­
in Fig. 2. It is noticed that orange juice consumption is positively lute Percentage Error (MAPE) are used as the error measure techniques.
correlated with grapefruit juice consumption. On the other hand, lemon RMSE can reduce the large error values into comparable metrics. On the
juice consumption is negatively correlated with both orange and other hand, MAPE focuses on the relative percentage amounts for
grapefruit juice consumption. This means the increased consumption of measuring the error that makes it easily interpretable. The method with
lemon juice in a quarter may decrease the orange and grapefruit juice minimum RMSE and MAPE is chosen. Then in the second stage, the
consumption for that quarter. Although the dataset does not contain qualitative criteria are analyzed. SWOT analysis using trapezoidal fuzzy
outliers, there are too many fluctuations in the data. In this research, numbers is carried out in this stage. Trapezoidal fuzzy numbers are used
these fluctuations are critically handled in the juice consumption fore­ to tackle the vagueness of human thought. The outputs of the first and
casting stage. second stages are used in the third stage, where the proposed multi-
objective model is developed to solve the SSOA planning problem. The
4. Methodology weighted-sum method is applied in this stage to solve the model.

Fig. 3 illustrates the proposed three-stage solution framework to

Fig. 2. Correlation among different citrus juice consumptions.

5
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

Fig. 3. The proposed solution framework for the SSOA planning.

4.1. Forecasting of demand u) are used to denote the non seasonal part, and the other 3 parameters
(i.e., d’ , r’ , andu’ ) are used to define the seasonal parts of the model
In this part, 3 forecasting methods are considered. The first method is (Dubey et al., 2021). The parameters d and d’ represent the data sta­
SARIMA, which is a well-known forecasting method. SARIMA is used as tionarity status. The Augmented Dickey-Fuller (ADF) test is performed
the base model of the performance evaluation benchmark in this study to check the data stationarity condition. In the null hypothesis of the
because it can handle seasonality, and its autoregression part can fore­ ADF test, it is assumed that the data is non-stationary, and the alterna­
cast data efficiently. Multilayer Perceptron (MLP) is used as the second tive hypothesis refers that the data is stationary (Gomez-Biscarri and
method because it is effective to solve nonlinear complex problems with Hualde, 2015). Table 3 includes the ADF test outcomes for three
good accuracy. In addition to these two methods, a modified Encoder- different types of juice consumptions in Ontario, Canada.
Decoder LSTM network is developed in this study to predict the demand. Considering the significance level as 0.05, the ADF tests’ p-values are
greater than the significance level. Hence, it is failed to reject the null-
4.1.1. Seasonal autoregressive integrated Moving average (SARIMA) hypothesis, and the ADF test results show there are seasonality/trend
method in the data. As the dataset in this research is a non-stationary dataset,
SARIMA method contains six parameters. 3 parameters (i.e., d, r, and differencing is performed to obtain data stationary. The parameters

6
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

Table 3 model for grapefruit juice. The model parameters (2, 1, 1) represent that
The results of data stationarity test. differencing is done two times to make the data stationary followed by
Type ADF test p- Null Alternative Stationary first order AR and MA. The ACF of residuals shows the correlation of
value hypothesis hypothesis status juice consumption among different quarters.
Orange juice 0.6215 TRUE FALSE Non- The Q-Q plot shows that the standard residuals on juice consump­
stationary tions are normally distributed. The Ljung-Box statistics reflects that the
Lemon juice 0.7429 TRUE FALSE Non- autocorrelation among the lags of different quarters is significant (i.e.,
stationary greater than the significance level of 0.05). The forecasting horizon for
Grapefruit 0.7518 TRUE FLASE Non-
juice stationary
the SARIMA model is set eight quarters. Forecasting results for the or­
ange juice, grapefruit juice, and lemon juice are displayed in Fig. 5. They
show a steady consumption growth on grapefruit juice. However, the
d and d’ in this SARIMA method show how many times differencing is forecasted demands for the orange and lemon juice have sudden spike
applied. After determining the number of required differencing, the on the fifth quarter. A 20 % of the forecasted demand values are
Auto Regression (AR) order is determined using the Partial Autocorre­ calculated using the forecasted demand because the company aims to
lation Function (PACF). AR is defined by using r and r’ for nonstationary fulfill that part of demand.
and stationary parts of the data, respectively. From PACF, the Auto­ The performance of the SARIMA model is presented in Table 4,
correlation Function (ACF) is determined. ACF is used to define the where h refers to the forecasting horizons, and λ refers to the number of
Moving Average (MA) order of the SARIMA method, denoted by u andu’ observations per year. It is observed that both error measures contain
(Carmona-Benítez and Nieto, 2020). Based on the ADF test, it is high values, which may cause poor forecasting accuracy. To achieve
observed that when the differencing is performed two times, the data better accuracy, two advanced machine learning methods are applied
becomes stationary. The value of the first SARIMA parameter is set to and discussed in next Sub-sections.
two. The ACF and PACF are also examined using the R package named
‘forecast’, and the values for both AR and MA are obtained. After 4.1.2. Multilayer Perceptron (MLP) method
determining all six parameters, the SARIMA is constructed as shown in To gain higher prediction accuracy, Multilayer Perceptron (MLP)
Eq. (1), where yt is the actual demand at time t, Ωt is the noise in the data method with backpropagation is applied on the juice dataset. The
at time t, ϖ denotes nonseasonal part, and ϖ ’ denotes the seasonal part developed MLP consists of three layers, namely the input layers, hidden
of the method. layers, and output layers (Wang et al., 2021). The input layers con­
taining multiple input nodes are denoted as X1 , X2 , ⋯, XN for N number
(ϖAR*ϖ ’ AR)(ϖ difference*ϖ ’ difference)* yt = (ϖMA*ϖ ’ MA)* Ωt (1) of inputs. Each input is associated with an assigned weight value W1 ,W2 ,
Fig. 4 illustrates the analysis of different parameters of the SARIMA ⋯,WN . The outputs of the input layers are used as an input to the hidden

Fig. 4. SARIMA parameters analysis on Grapefruit juice consumption.

7
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

Fig. 5. The forecasted demand for citrus juice in Ontario using the SARIMA method.


Table 4 ̂N =
Y HN WN + Δy (4)
The SARIMA model performance: h = 8 (d = 2, r = 1, u = 1) (d’ = 2,r’ = 1,u’ = N
1) λ = 4.
1
Error measure Orange juice Grapefruit juice Lemon juice σ N(Y) = (5)
1 + exp− ̂Y N
RMSE 1,885.14 1,131.52 1,884.61
MAPE 5.62 46.75 184.40 In the next step, the errors (εN ) are calculated for each output node by
deducting the produced Y ̂ N of the output node from the actual YN as
layers, represented by H1 ,H2 ,⋯,HN , as shown in Eq. (2). Bias values are presented in Eq. (6). Then, the total error is calculated by summing up
added in the hidden layers and the output layers, symbolized using ΔH the error produced by each output node as represented in Eq. (7).
and Δy , respectively. The hidden layers contain an activation function. 1
The Sigmoid activation function (Liu et al., 2020) is used in this devel­ εN = ̂N)
(YN − Y (6)
N
oped method as shown in Eq. (3). σN(H) represents the activation func­

tion, and exp stands for exponent. If the node is activated, it sends the εTotal = εN (7)
value to the next layer. This process is repeated through the output N

layers. The outputs of the hidden layers (HN ) are used as the inputs of the To reduce the error (εN ), the backpropagation method is used in each
̂ N ), as shown in Eq. (4). The activation function for the
output layer ( Y perceptron. The weight values are updated in a fashion that aims to
output layers is shown in Eq. (5), and it is represented by σ N(Y) . minimize the gap between the produced Y ̂ N and the actual YN . Error
∑ gradients (∇ε ) with respect to the weights (WN ) are calculated to see
HN = XN W N + Δ H (2)
N
how much change in the weights can affect the total error (εTotal ). To do
so, the partial derivative of εTotal with respect to WN is performed as given
1 in Eq. (8).
σ N(H) = HN
(3)
1 + exp− ∂εTotal
∇ε = (8)
∂WN

8
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

The weighted values are updated based on the minimum values of Table 5
∇ε . For each type of juice, a perceptron is created to predict the demand The performance of the MLP method.
for a specific quarter. All perceptrons are processed in parallel. The Error measure Orange juice Grapefruit juice Lemon juice
prediction results are shown in Fig. 6. The implementation of back­
RMSE 991.75 523.45 951.28
propagation helps reduce the prediction error during the training phase MAPE 3.15 12.25 36.33
as depicted in Fig. 6.
The performance measures of the MLP method are shown in Table 5.
It is observed that MLP can reduce the RMSE by approximately 50 % in technique, proposed by (Singh and Singh, 2020). As shown in Eq. (10),
comparison to the SARIMA. the normalized data is denoted using Γ. ϋ is the data value, μ is the
average of the dataset, and D is the standard deviation of the dataset.
4.1.3. Long-Short Term Memory (LSTM) network ( ) ϋ − μ
Instead of implementing traditional LSTM network, a modified Long- Γ gfluctuation = (10)
D
Short Term Memory (LSTM) network is proposed in this study to reduce
the forecasting errors. In this modified approach, three LSTM networks In the next step, the normalized fluctuated data is passed through the
are used in parallel. Despite of passing the whole data as an input of the third LSTM network. All three autoencoder-decoder LSTM networks
LSTM network, the data is decomposed first to separate the trend (ɠ trend ) forecast the trend (̂ y ɠtrend ), seasonality (̂
y ɠseasonal ), and normalized fluctu­
of the data from the data seasonality (ɠ seasonal ) and fluctuation ation (̂y Γ(ɠ ) ) separately,
fluctuation
but parallelly for the next eight quarters.
(ɠ fluctuation ). An additional normalizer is used to normalize the fluctua­ The forecasted components are added to get the actual forecast as shown
tions values. The trend, seasonality, and the normalized fluctuations are in Eq. (11). The proposed multivariate multistep LSTM network archi­
then passed through three parallel LSTM networks. Fig. 7 is presented as tecture is presented in Fig. 8.
an example, where the trend, seasonality, and fluctuation are separated
from the actual data of orange juice consumptions for the 60 years y Ò¦orange = ̂y ɠtrend + ̂y ɠseasonal + ̂
̂ y Γ(ɠfluctuation ) (11)
timespan.
Table 6 contains the forecasted decomposed parameters for three
Fig. 7 displays that the trend is independent of the seasonality. So,
different juices demands. The total forecasted demand is the summation
the orange juice data components are additive. Hence, the train data
of those three parameters values. It is noticed that the forecasted fluc­
Ò¦orange , can be written as in Eq. (9). tuations have very less impacts on the data because of the applied
Ò¦orange = ɠ trend + ɠ seasonal + ɠ fluctuation (9) normalization technique. However, the forecasted seasonal parameters
have high impacts on the total forecasted values. The model is imple­
In this proposed method, the trend and seasonal data are passed mented using Python programming in Keras framework on the top of
through two separate LSTM networks. The desired prediction horizon is Tensorflow environment. The Rectified Linear Activation Function
for eight quarters. Hence, the LSTM network is designed to forecast (ReLU) (Schmidt-Hieber, 2020) is used as the activation function in this
multistep forecasting. To perform multivariate and multistep pre­ proposed LSTM network. The data combiner is added at the end which
dictions, the autoencoder decoder LSTM method is used. The fluctua­ combines the outputs to get the final forecast.
tions components are normalized using Z-Score normalization Fig. 9 shows the error distribution of three forecasting models

Fig. 6. The forecasted demand for citrus juice in Ontario using the MLP method.

9
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

Fig. 7. Training data decomposition for orange juice. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of
this article.)

Fig. 8. The proposed multivariate multistep LSTM architecture.

considered in Stage 1. The error values are for orange juice forecasting. In this research, different qualitative criteria are considered. The
In this figure, the median error for MLP and SARIMA are above 1,500, decision-making body includes three Decision-Makers (DMs) with
whereas the median error for the modified LSTM is below 900. It is also different degrees of expertise. Thus, DM = 3. The qualitative criteria are
observed that the spread of the first quartile is minimum for the LSTM. judged by these DMs. The expertise of them is vital to evaluate the SS
This means that LSTM produces fewer errors during forecasting than the criteria. Hence, their years of expertise are considered, using a linguistic
other two models. variable as shown in Fig. 10. The weighted expertise can be interpreted
as low level (0, 0, 2, 4), moderate level (0, 2, 6, 8), and high level (4, 6, 8,
4.2. SWOT analysis for the qualitative factors 8). First, the qualitative criteria are determined by the DMs. The internal
factors include quality, service stability, waste management, and sup­
Prioritizing suppliers based on qualitative SS criteria is difficult as it plier’s past performance. On the other hand, the external factors
involves human judgements. These judgements might appear to have a comprise the use of Greenhouse Gas (GHG) legislation, suppliers’ health
lack of clarity and are sometimes hard to interpret. The Strengths, and safety policies, and suppliers’ market reputation. Then, the
Weaknesses, Opportunities, and Threats (SWOT) analysis has been following steps are carried out to translate those criteria and to deter­
adopted by some researchers (e.g., Sanito et al., 2020; Büyüközkan et al., mine the weights of the suppliers.
2021; Hosseini et al., 2021; Qaiser, 2022) across different industries for Step 1: In this step, the DMs assess the qualitative criteria (both in­
qualitative factors analysis. In this part, a combined method of Trape­ ternal and external criteria) for the available suppliers. These assess­
zoidal Fuzzy Numbers (TrFN) and SWOT analysis is developed to ments are converted into the numeric weights using a linguistic scale. In
translate the SS qualitative criteria and to determine the weights of the this study, a linguistic scale with five variables, namely Low (L), Medium
suppliers. The TrFN is used to tackle the vagueness of human thoughts. Low (ML), Medium (M), Medium High (MH), and High (H) is used.
SWOT analysis is used in this study because it is a unique technique that These linguistic variables are quantified using TrFN. Fig. 11 shows how
can consider external factors (e.g., opportunities) in addition to internal the weights are assigned according to the different levels of linguistic
factors. variables, using trapezoidal fuzzy numbers. This figure also illustrates

10
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

Lemon
20 % of the forecasted demands

790.2

833.6

848.2

839.8

826.0

843.8

850.0

848.0
Grapefruit

205.3

212.4

225.5

207.8

212.6

206.8

216.0

220.2
5,699.7

6,436.8

6,562.4

6,621.9

6,646.2

7,155.6

7,293.8

7,153.8
Orange

3,951.0

4,168.0

4,241.0

4,199.0

4,130.0

4,219.0

4,250.0

4,240.0
Lemon
Grapefruit

1,026.7

1,062.0

1,127.5

1,039.0

1,063.0

1,034.0

1,080.0

1,101.0
Forecasted demands

Fig. 9. Forecasting models comparison during forecasting of orange juice. (For


interpretation of the references to colour in this figure legend, the reader is
referred to the web version of this article.)
28,498.3

32,184.0

32,812.0

33,109.4

33,230.8

35,778.2

36,469.0

35,769.0
Orange

the interpretation of the linguistic scale. Once a DM provides the eval­


uation using any of the five variables (L, ML, M, MH, H) used in the
linguistic scale, the trapezoidal fuzzy numbers of the corresponding
Encoder-Decoder LSTM forecast for lemon

variable are multiplied first with the fuzzy numbers of the corresponding
Seasonality

DM’s expertise. The evaluations of all DMs are quantified in the similar
manner. These quantified numbers are summed up for each type of fruit,
− 48.5

49.2

15.8

15.0

− 13.3
4.2

22.6
− 6.7

and then are divided by the number of DMs. In this study, the trape­
zoidal fuzzy numbers are aggregated to get the weighted criteria (Wspj )
Fluctuation

using Eq. (12). In this formula, s represents suppliers, p represents fruits,


and j represents criterion.
–23.4

113.1
51.4

–23.4

113.1
58.9

91.4

88.9

Wspj1 + Wspj2 + ⋯ + WspjDM


Wspj = (12)
DM
4,022.9

4,067.4

4,112.1

4,135.9

4,138.4

4,140.9

4,143.6

4,128.5
Trend
juice

Seven qualitative criteria are used in this study. The weight assign­
ment procedure for ‘Quality’ criterion is shown in Table 7, as an
example. There are five available suppliers in this study. To defuzzify the
Encoder-Decoder LSTM forecast for grapefruit

trapezoidal fuzzy numbers, the centroid method is applied. The defuz­


zification of the trapezoidal fuzzy number ℘ = {f1 , f2 , f3 , f4 } is calculated
Seasonality

using Eq. (13). Table 8 includes the defuzzified weights of the qualitative
− 62.2

10.3

95.7

− 10.2

22.1
− 3.9

9.7
–22.1

criteria.
1
Defuzzified(℘) = (f1 + f2 + f3 + f4 ) (13)
Fluctuation

4
Step 2: The importance levels of the criteria are not the same, and it is
− 5.6

− 1.1

− 5.6

− 1.1
26.0

− 19.3

26.0

− 19.3
The forecasted demand of citrus juice using Encoder-Decoder LSTM network.

important to understand the importance level of each criterion before


deciding. In this step, the DMs assign the importance level of each
1,062.9

1,057.4

1,051.0

1,044.0

1,047.3

1,061.8

1,077.1

1,092.4

qualitative criterion and calculate the weights by using the linguistic


Trend
juice

variables. Defuzzification of those weights are obtained using Eq. (13).


Then, the deffuzified weights are normalized using Eq. (14) to unify the
scales of the key factors. Ψspj represents the normalized weights of cri­
Encoder-Decoder LSTM forecast for orange

terion j, for supplier s, and fruit p. Table 9 contains the results.


Seasonality

− 1,816.2

615.7

384.1

− 390.3

− 1,143.0

790.2

369.5

− 959.2

Wspj
Ψspj = ∑ (14)
Max s Wpj

Step 3: In this step, the defuzzified weights of suppliers from Table 8


Fluctuation

are normalized using Eq. (15). In this formula, ℵspj represents the
114.4

− 130.6

114.4
76.9

− 126.9

130.6
93.2

− 93.2

normalized weights of supplier s for fruit p and criterion j. Besides, ℘spj


denotes the defuzzified weights of a supplier. Then, the normalized
weights of criteria from Table 9 are multiplied with the corresponding
30,200.1

31,491.4

32,558.5

33,406.5

34,259.4

35,114.9

35,968.9

36,821.4

normalized weights of suppliers. These multiplied values are added


Trend
juice

together based on the type of fruit. As an example, the normalized


weights of Supplier 1 for orange are multiplied with the corresponding
Year and Quarter

normalized weights of each internal criterion. Then, these multiplied


Jan-Mar

Jan-Mar
Apr-Jun

Apr-Jun
Oct-Dec

Oct-Dec
Jul-Sep

Jul-Sep

values for each internal criterion are summed up to get a single value for
orange. The similar procedure is applied for the external criteria.
Table 6

Q1:

Q2:

Q3:

Q4:

Q1:

Q2:

Q3:

Q4:

Finally, the calculated values are normalized in a way that the sum­
2022

2022

2022

2022

2023

2023

2023

2023

mation of all suppliers’ weights for any specific fruit is equal to one. Eq.

11
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

Fig. 10. Identifying the weights of the decision-makers.

Fig. 11. Quantifying the linguistic variables.

Table 7 Table 8
Weight calculation of Quality. Defuzzified weights of qualitative criteria.
Supplier Product DM1 DM2 DM3 Weight Defuzzification Internal criteria Product s=1 s=2 s=3 s=4 s=5
(s)
Quality Orange 20.35 16.00 32.65 27.35 26.35
1 Orange M M M (4, 10.7, 20.35 Grape 21.35 15.00 32.65 27.35 10.00
26.7, 40) Lemon 17.00 28.35 25.68 27.33 26.35
Grape MH M M (4, 10.7, 28, 21.35 Service stability Orange 24.65 15.00 32.65 27.35 27.33
42.7) Grape 22.00 16.00 32.65 27.35 23.70
Lemon MH M ML (1.3, 6.7, 17.00 Lemon 17.68 28.35 25.68 27.33 27.33
22.7, 37.3) Waste management Orange 23.70 12.33 34.00 26.35 8.33
2 Orange M M ML (1.3, 6.7, 16.00 Grape 27.35 16.00 29.35 27.35 11.03
21.3, 34.7) Lemon 19.33 28.35 29.33 27.33 8.33
Grape ML M ML (1.3, 6.7, 20, 15.00 Supplier’s past Orange 24.00 13.30 27.35 20.35 13.30
32) performance Grape 27.35 16.00 27.35 27.35 19.33
Lemon ML H MH (6.7, 17.3, 28.35 Lemon 21.35 28.35 29.35 26.65 9.33
38.7, 50.7) External criteria Product s=1 s=2 s=3 s=4 s=5
3 Orange H MH H (9.3, 20, 44, 32.65 GHG legislations Orange 16.00 9.33 22.00 17.68 21.35
57.3) Grape 14.68 19.68 18.65 13.68 14.33
Grape H MH H (9.3, 20, 44, 32.65 Lemon 13.00 25.00 21.35 11.00 9.33
57.3) Health and safety Orange 26.35 28.33 34.00 29.03 31.00
Lemon MH M MH (6.7, 14.7, 25.68 Grape 29.03 26.65 34.00 25.68 32.65
33.3, 48) Lemon 22.00 30.33 30.00 27.33 34.00
4 Orange M MH MH (6.7, 16, 36, 27.35 Market reputation Orange 32.65 28.33 34.65 32.65 30.33
50.7) Grape 26.00 31.00 31.00 21.35 24.00
Grape M MH MH (6.7, 16, 36, 27.35 Lemon 29.03 34.00 29.03 29.03 29.03
50.7)
Lemon ML M H (9.3, 18.7, 27.33
36, 45.3)
5 Orange L M H (9.3, 18.7, 26.35 ℘spj
ℵspj = (15)
34.7, 42.7) Max℘spj
Grape L M L (0, 2.7, 13.3, 10.00
24) ∑∑
j Ψspj *ℵspj
Lemon L M H (9.3, 18.7, 26.35 (16)
p
Isp = ∑ ∑ ∑
34.7, 42.7) s p j Ψspj *ℵspj

(16) shows the suppliers’ total internal weighted values. In this equa­ 4.3. Multi-objective optimization model for SSOA planning
tion, Isp represents the total weighted values for the internal criteria, and
Ψ represents the normalized weights. A similar procedure is applied to In this part, a Multi-objective Mixed-Integer Linear Programming
get the total external weighted values (Esp ) for the suppliers. The results (MOMILP) optimization model is developed to identify the suitable
of this step are written in Table 10. The calculated values of Isp and Esp suppliers for different fruits, and to decide the order amounts from them.
from Table 10 are used in the multi-objective optimization model. As the demand values for juice (in KL) refer to the required amounts of
fruits (in tonnes), the forecasted demand values from Sub-section 4.1 are

12
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

Table 9
Importance of criteria.
Criteria DM1 DM2 DM3 Weight Defuzzified weight Normalized weight

Quality (internal) H MH H (9.3, 20, 44, 57.3) 32.65 0.26


Service stability (internal) H MH M (4, 12, 33.3, 49.3) 24.65 0.19
Waste management (internal) ML MH H (9.3, 20, 40, 50.7) 30.00 0.27
Supplier’s past performance (internal) M ML MH (6.7, 13.3, 28, 40) 22.00 0.19
GHG legislations (external) L ML L (0, 1.3, 9.3, 18.7) 7.33 0.08
Health and safety (external) H M MH (6.7, 14.7, 34.7, 49.3) 26.35 0.18
Market reputation (external) L M H (9.3, 18.7, 34.7, 42.7) 26.35 0.17

Table 10
Normalized scores for suppliers, products, and criteria.
Criteria Normalized weights of criteria Product Defuzzified weights Normalized weights

s=1 s=2 s=3 s=4 s=5 s=1 s=2 s=3 s=4 s=5

Quality 0.26 Orange 20.35 16.00 32.65 27.35 26.35 0.62 0.49 1.00 0.84 0.81
Grape 21.35 15.00 32.65 27.35 10.00 0.65 0.46 1.00 0.84 0.31
Lemon 17.00 28.35 25.68 27.33 26.35 0.60 1.00 0.91 0.96 0.93
Service stability 0.19 Orange 24.65 15.00 32.65 27.35 27.33 0.75 0.46 1.00 0.84 0.84
Grape 22.00 16.00 32.65 27.35 23.70 0.67 0.49 1.00 0.84 0.73
Lemon 17.68 28.35 25.68 27.33 27.33 0.62 1.00 0.91 0.96 0.96
Waste management 0.27 Orange 23.70 12.33 34.00 26.35 8.33 0.70 0.36 1.00 0.78 0.24
Grape 27.35 16.00 29.35 27.35 11.03 0.93 0.55 1.00 0.93 0.38
Lemon 19.33 28.35 29.33 27.33 8.33 0.66 0.97 1.00 0.93 0.28
Supplier’s past performance 0.19 Orange 24.00 13.30 27.35 20.35 13.30 0.88 0.49 1.00 0.74 0.49
Grape 27.35 16.00 27.35 27.35 19.33 1.00 0.59 1.00 1.00 0.71
Lemon 21.35 28.35 29.35 26.65 9.33 0.73 0.97 1.00 0.91 0.32
Total weighted values for internal criteria (Isp ) Orange 0.20 0.13 0.28 0.23 0.16
Grape 0.22 0.14 0.27 0.24 0.13
Lemon 0.16 0.24 0.23 0.23 0.15
GHG legislations 0.08 Orange 16.00 9.33 22.00 17.68 21.35 0.73 0.42 1.00 0.80 0.97
Grape 14.68 19.68 18.65 13.68 14.33 0.75 1.00 0.95 0.70 0.73
Lemon 13.00 25.00 21.35 11.00 9.33 0.52 1.00 0.85 0.44 0.37
Health and safety 0.18 Orange 26.35 28.33 34.00 29.03 31.00 0.78 0.83 1.00 0.85 0.91
Grape 29.03 26.65 34.00 25.68 32.65 0.85 0.78 1.00 0.76 0.96
Lemon 22.00 30.33 30.00 27.33 34.00 0.65 0.89 0.88 0.80 1.00
Market reputation 0.17 Orange 32.65 28.33 34.65 32.65 30.33 0.94 0.82 1.00 0.94 0.88
Grape 26.00 31.00 31.00 21.35 24.00 0.84 1.00 1.00 0.69 0.77
Lemon 29.03 34.00 29.03 29.03 29.03 0.85 1.00 0.85 0.85 0.85
Total weighted values for external criteria (Esp ) Orange 0.19 0.17 0.23 0.20 0.21
Grape 0.19 0.21 0.23 0.17 0.20
Lemon 0.17 0.23 0.21 0.18 0.20

directly used in the mathematical model. Suppliers’ weights for different Gspq : Capacity of supplier s for fruit p in period q (in tonnes).
qualitative criteria along with the weights of internal and external Decision Variables:
criteria from Sub-section 4.2 are also used in this part. The ordering cost Ospq : Order amount, fruit p from supplier s in period q.
of the manufacturing company is constant, however, the cost is high (in Vpq : Inventory level, fruit p and period q.
thousand dollars), as it contains inspection and approval fees of fruit Bspq : Binary (0 or 1) variable. When fruit p is obtained from supplier s
farm inspectors. The details are as follows: & in period q, it is set to one. If not, it is zero.
Sets: ∑∑∑ ∑∑∑ ∑∑
Fruits: P = {1…p…P}. Min δ1 = Cspq Ospq + Ksp Bspq + Jpq Vpq (17)
Periods: Q = {1…q…Q}.
s p q s p q p q

Suppliers: S = {1…s…S}. ∑∑∑( )


Parameters: Max δ2 = β1 Isp + β2 Esp Ospq
Isp : Total weighted values for internal criteria, supplier s, fruit p.
s p q

Esp : Total weighted values for external criteria, supplier s, fruit p. (18)
β1 : Weight of the internal criteria. ∑∑∑
Min δ3 = Fspq Ospq (19)
β2 : Weight of the external criteria. s p q
Ksp : Ordering cost of fruit p from supplier s (in thousand dollars).
∑∑∑
Cspq : Purchasing cost of fruit p from supplier s in period q (in thou­ Max δ4 = Rspq Ospq (20)
sand dollars). s p q
Jpq : Holding cost for fruit p in period q (in thousand dollars).
∑∑∑
Rspq : On-time delivery rate, supplier s for fruit p in period q (in Min δ5 = Aspq Ospq (21)
percentage). s p q

Aspq : Defect rate of fruit p, supplier s in period q (in percentage). Subject to:
Fspq : Carbon footprint values of fruit p, supplier s in period q (in ∑
tonnes). Ospq = Dpq ∀p, q (22)
Dpq : Demand of fruit p in period q (in tonnes). s

13
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

∑ Table 11
Ospq + Vp(q− − Dpq = Vpq ∀p, q (23)
1)
The data for the developed multi-objective optimization model.
s
C11q = $2.9 (q = 1, 4, 5, 8) C12q = $4.3 (q G11q = 3,400 T; G12q = 120 T
Ospq ≤ Gspq Bspq ∀s, p, q (24) Cs1q = $3.1 (s = 1, 2, 4) = 1, 5) G13q = 500 T
Cs1q = $3.3 (s = 1, 2, 5; q = C12q = $4.9 (q G21q = 1,700 T; G22q = 100 T
2, 3, 6, 7) = 2, 6) G23q = 400 T
Vpq ≥ 0; Ospq ≥ 0∀s, p, q (25)
Cs2q = $5.4 (s = 1, 3; q = 2, C12q = $4.6 (q G31q = 2,300 T; G32q = 140 T
3, 6, 7) = 4, 8) G33q = 300 T
Bspq = 0or1∀s, p, q (26) Cs3q = $1.9 (s = 1, 4, 5) C13q = $1.8 (q G41q = 3,200 T; G42q = 180 T
Cs3q = $2.1 (s = 1, 2; q = 1, = 1, 5) G43q = 600 T
The proposed optimization model consists of five objective functions 3, 4, 5, 7, 8) C22q = $5.1 (q G51q = 3,100 T; G52q = 210 T
(δ1 , δ2 , δ3 , δ4 , andδ5 ). The first objective function (δ1 ) minimizes the total Cs2q = $4.5 (s = 2, 4; q = 1, = 2, 6) G53q = 500 T
cost. The purchasing costs of the fruits, holding costs, and ordering costs 4, 5, 8) C22q = $5.7 (q Fspq = 0.13 tonnes (s = 1, 2, 3,
are cogitated as the total cost in Eq. (17). In addition, the weights of the Cs2q = $4.8 (s = 2, 3, 5; q = = 3, 7) 4; q = 1, 4, 7)
1, 4, 5, 8) C21q = $3.5 (q Fspq = 0.12 tonnes (s = 1, 2, 3,
suppliers obtained from the SWOT analysis are maximized using Eq.
Cs3q = $2.2 (s = 2, 5; q = 2, = 3, 7) 4; q = 2, 5, 8)
(18). There are four internal and three external criteria considered in the 3, 6, 7) C31q = $3.7 (q Fspq = 0.11 tonnes (s = 1, 2, 3,
SWOT analysis. Thus, the values of β1 and β2 are assigned as β1 = 4/7 Cs3q = $2.4 (s = 2, 3; q = 2, = 2, 6) 4; q = 3, 6)
and β2 = 3/7. The total weight is equal to one (β1 + β2 = 1). Eq. (19) 3, 6, 7) C31q = $3.8 (q F5pq = 0.19 tonnes (q = 1, 4, 7)
Cs1q = $3.4 (s = 3, 5; q = 1, = 3, 7) F5pq = 0.17 tonnes (q = 2, 5, 8)
minimizes the carbon footprint of the suppliers. The rates of on-time
3, 4, 5, 7, 8) C32q = $5.9 (q F5pq = 0.15 tonnes (q = 3, 6)
delivery from the suppliers are maximized using Eq. (20). To ensure Cs1q = $3 (s = 4, 5; q = 1, 2, = 3, 7) Isp and Esp are obtained from
the quality of the fruits, the defect rates for suppliers are minimized in 4, 5, 6, 8) C33q = $2.6 (q Stage 2.
Eq. (21). Eq. (22) is used as a constraint to make sure that the demand is C43q = $1.6 (q = 1, 4, 5, 8) = 3, 7)
fulfilled. Another constraint in Eq. (23) bounds the number of orders by Cs2q = $5.2 (s = 3, 4; q = 3, C42q = $4.1 (q
4, 7, 8) = 1, 5)
considering the current inventory levels and the forecasted demands.
C33q = $2.3 (q = 1, 4, 5, 8) C42q = $4.7 (q
Constraint (24) confirms that the order quantities do not exceed the C41q = $2.7 (q = 1, 4, 5, 8) = 2, 6)
suppliers’ capacities. The constraints of non-negative and binary vari­ Dpq are obtained from Stage C43q = $1.7 (q
ables are shown in Eq. (25) and Eq. (26), respectively. 1. = 2, 6)
R1pq = 0.85; R2pq = 0.77; C52q = $4.4 (q
R3pq = 0.88 = 1, 5)
4.4. Weighted-sum method R4pq = 0.91; R5pq = 0.89 C52q = $5 (q =
A1pq = 0.15; A2pq = 0.12; 2, 6)
A3pq = 0.19 C52q = $5.5 (q
The solutions of a multi-objective model are called efficient solu­ A4pq = 0.11; A5pq = 0.14 = 3, 7)
tions, instead of one unique optimal solution (Chagas and Wagner, C53q = $2 (q =
2022). In this part, the weighted-sum method is applied for solving the 2, 6)
Ksp = $4
multi-objective optimization model to obtain a set of efficient solutions.
Jpq = $1
In this technique, all objective functions are merged as a new single
objective function. During this merging process, weights are assigned to
each objective function in a way that the total weight becomes one
(Aurzada, 2020; Ouadfel and Abd Elaziz, 2021; Chagas and Wagner, for this developed model is presented in Table 12. A set of nine efficient
2022). In this method, usually the weights are assigned by the decision- solutions are produced by changing the assigned weights for different
makers and can be tuned into different numbers to obtain a set of effi­ objectives. These efficient solutions are provided to the decision-makers.
cient solutions (Amin and Zhang, 2013; Jauhar et al., 2021; Mishra
et al., 2021). The scalarized objective (δ6 ) is minimized in this study. Eq. 5. Discussions
(27) shows the formula where five objectives are denoted, using δ1 ,δ2 ,δ3 ,
δ4 , andδ5 . Five weights (w1 , w2 , w3 , w4 , andw5 ) are multiplied to the cor­ Based on the data from the production section of our studied com­
responding objectives. The weights are defined in a way such that the pany, it is noticed that the ratio of produced raw juice and required

summation of all weights is one (i.e., m wm = 1, if m is the index of amounts of fruits is almost 1:2 (e.g., one liter of raw orange juice quires
objective function). Besides, the assigned weights are non-negative real almost two kilograms of oranges). Besides, during the manufacturing
numbers (i.e., wm ≥ 0). process, the ratio of raw juice and water is almost 50:50 (e.g., one liter of
manufactured orange juice contains 500 ml of raw orange juice). Hence,
Minδ6 = w1 δ1 − w2 δ2 + w3 δ3 − w4 δ4 + w5 δ5 (27)
the ratio of manufactured juice and the required amount of fruits be­
Quantitative parameters such as production capacities, unit costs, comes 1:1 (e.g., one liter of manufactured orange juice requires one
ordering costs, inventory costs, and other relevant information are kilogram of orange). For this reason, the forecasted demand for different
collected from the Ontario Fruits and Vegetables Growers’ Association manufactured juice is directly used as the demand of the fruit for our
(OFVGA Directory, 2021). Table 11 includes the data used in the opti­ manufacturing company in the developed MOMILP model. This math­
mization model. The literature and Stages 1 and 2 of this paper are some ematical model intends to solve multiple objectives to determine order
of the sources of the data. The websites of the suppliers are also used to allocations from the selected suppliers. The units of different parameters
fetch the rest of the data. The fruits’ price (in thousand dollars) varies for are also not the same (e.g., purchasing cost in thousand dollars, sup­
different suppliers and quarters. The carbon footprint also fluctuates pliers’ capacity in Tonnes, and on-time delivery in percentage). Hence,
from 0.11 tonnes to 0.19 tonnes per thousand Kiloliters (KL) of juice the MOMILP model is used to solve the SSOA problem. The objectives of
products for different suppliers. Besides, different suppliers have various this model may have conflict. For example, the purchasing cost from one
production capacities for different fruits, ranging from 100 tonnes (T) to supplier can be less than the other suppliers, but the on-time delivery
3,400 tonnes (T) (Miranda-Ackerman et al., 2017). can be worse than the other suppliers. Moreover, it is multi-objective
The weighted-sum technique is simple to be implemented, and it is because they have different units (e.g., cost ($), on-time delivery
excellent to address a convex set (Chen and Sung, 2022). The developed (percentage)).
multi-objective optimization model is solved using the GAMS Studio To analyze the trade-offs among the five different objective func­
software, version 1.5.4. This software can solve optimization models, tions, the value path approach is applied in this study. The value path
efficaciously (Chagas and Wagner, 2022). The set of efficient solutions approach provides a graphical relevance for decision-making by

14
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

Table 12
The MOMILP models’ efficient solutions using the weighted-sum technique.
w1 w2 w3 w4 w5 δ1 δ2 δ3 δ4 δ5

0.20 0.20 0.20 0.20 0.20 179,330.44 12,711.49 7,611.16 54,658.11 7,998.67
0.025 0.575 0.025 0.075 0.3 189,414.95 14,065.14 7,503.59 54,982.97 8,710.14
0.1 0.025 0.025 0.75 0.1 182,308.36 12,350.70 8,847.93 55,627.18 7,756.40
0.037 0.063 0.5 0.3 0.1 180,078.96 12,859.71 7,503.59 54,629.48 8,094.71
0.375 0.2 0.075 0.05 0.3 179,330.44 12,712.93 7,625.54 54,667.70 7,996.27
0.19 0.4 0.01 0.2 0.2 179,330.44 12,712.93 7,625.54 54,667.70 7,996.27
0.05 0.375 0.037 0.525 0.063 190,255.29 13,787.96 8,081.93 55,591.33 8,652.03
0.2 0.01 0.19 0.2 0.4 179,330.44 12,712.93 7,625.54 54,667.70 7,996.27
0.4 0.3 0.1 0.1 0.1 179,330.44 12,712.93 7,625.54 54,667.70 7,996.27

illustrating the trends for the various objectives (Fernandes et al., 2021). Hence the machine learning models are chosen in a way so that they can
This approach has been applied in some recent related research (e.g., handle large data. It is proven by related research that SARIMA, MLP,
(Fonseca et al., 2021)). The efficient solutions obtained from the and LSTM are useful for large datasets (Chaturvedi et al., 2022; Yin
developed optimization model are normalized first, as written in et al., 2022). It is noticed that the developed LSTM network decreases
Table 13. The developed optimization model comprises five objectives, the average total error by 71 % and 68.5 % in comparison with the
where objectives one, three, and five are for minimization. On the other SARIMA and MLP models, respectively. The values of error metrics can
hand, objectives two and four are for maximization. For the normali­ be changed based on the fruits data. Considering the LSTM error metrics
zation process, the minimization objectives are multiplied first by a as an example, the RMSE of orange is higher than the grapefruit,
negative one. Then, the objective values of the efficient solutions are whereas the MAPE of orange is lower than the grapefruit. RMSE is
divided by the minimum value of the corresponding objective. After calculated based on the root-mean error of the data, and it has been
performing the division, each value of the objectives is multiplied with widely adopted in the forecasting research (Karunasingha, 2022). In this
the corresponding weight to get the normalized scale. The normalized study, orange consumption values are much higher than the grapefruit
values are plotted to identify the trends for different objectives, as consumption values. So, the forecasted orange values are also higher
depicted in Fig. 12. The intersections in this figure confirm that the than the forecasted grapefruit values. As the root-mean error of larger
obtained solutions are efficient solutions. values is higher than the root-mean error of the smaller values (Zhang
The developed model is compared to a similar model where the et al., 2022), the RMSE of orange is higher than the grapefruit RMSE. On
values from Stage 1 are not fetched. Instead, the values from the last the other hand, MAPE is not affected by the large forecasting values, as it
known historical data for demand parameter is used in the new model. considers the average percentage of absolute error (Liu et al., 2022).
The goal is to observe the differences in the model’s outputs for supplier Therefore, MAPE is considered by the researchers where the data values
selection and related orders from them. As an example, for the orange are high (Kolassa, 2020). In this study, we have both high and low values
demand in Quarter 1, Suppliers 1 and 4 are selected when the first stage for different fruits. Hence, both RMSE and MAPE are used, and then the
is included in the proposed SSOA framework. On the other hand, Sup­ average is calculated to get the average total error, as written in
pliers 1, 4, and 5 are chosen when the first stage is not considered in the Table 15.
SSOA framework. The order amounts also vary, depending on the Although, the LSTM network is widely used in recent studies, in this
integration of the first stage with the framework. As shown in Fig. 13, research instead of traditional LSTM, an ensemble of three LSTM is used
the order quantity from Supplier 1 is 2,499.7 Tonnes with the integra­ to forecast three different components of our studied dataset. The
tion of the first stage. However, the quantity increases to 3,400 Tonnes dataset is decomposed into trend, seasonality, and fluctuation compo­
when the first stage is not considered. The results show that the first nents. Desired forecasting accuracy may not be achieved if there are
stage affects both supplier selection and order allocation. fluctuations in the data. As reported in Section 3 (problem statement),
In this part, the effects of variability in the weights of the internal and the studied data contains irregular fluctuations. Thus, we separated the
external criteria on the selected suppliers and the corresponding orders fluctuation components from the actual data. Then, this data component
are investigated. Table 14 includes the variations in the selected sup­ is normalized using a popular and powerful normalization (Z-Score)
pliers for different values of β1 andβ2 for Lemon juice in Quarter 4. It is technique. When the irregular fluctuation component is normalized, all
observed that Suppliers 4 and 5 are selected, when β1 < 0.5andβ2 > 0.5. three data components (trend, seasonality, and fluctuations) are fore­
However, the selected suppliers are changed to 1 and 5 for casted separately, using three LSTM in parallel. The forecasted data el­
β1 ≥ 0.5andβ2 ≤ 0.5. These results show that the values of β1 andβ2 have ements are combined to get the actual forecast. To the best of our
impacts on both the supplier selection and the related orders. knowledge, this approach is novel where the data components are
The performance comparison results among the developed LSTM separated, fluctuations are normalized using Z-score normalization,
network, MLP, and SARIMA models are discussed in Table 15. The then each component is predicted separately, and then components are
studied dataset contains different juice consumption records from 1960. aggregated for each step to get better forecasting accuracy.

Table 13
The normalized efficient solutions for the MOMILP model.
w1 w2 w3 w4 w5 δ1 δ2 δ3 δ4 δ5

0.20 0.20 0.20 0.20 0.20 0.18852 0.20584 0.17204 0.20010 0.18366
0.025 0.575 0.025 0.075 0.3 0.02489 0.65482 0.02120 0.07549 0.30000
0.1 0.025 0.025 0.75 0.1 0.09582 0.02500 0.02500 0.76370 0.08905
0.037 0.063 0.5 0.3 0.1 0.03502 0.06560 0.42403 0.30000 0.09293
0.375 0.2 0.075 0.05 0.3 0.35347 0.20587 0.06464 0.05003 0.27541
0.19 0.4 0.01 0.2 0.2 0.17909 0.41173 0.00862 0.20014 0.18361
0.05 0.375 0.037 0.525 0.063 0.05000 0.41864 0.03380 0.53424 0.06258
0.2 0.01 0.19 0.2 0.4 0.18852 0.01029 0.16375 0.20014 0.36722
0.4 0.3 0.1 0.1 0.1 0.37703 0.30880 0.08618 0.10007 0.09180

15
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

Fig. 12. Visualizing the efficient solutions using the value path approach.

Lastly, the impacts of the forecasting models on the selected sup­


pliers and the orders are discussed. Table 16 presents the impacts of
these three forecasting models during the supplier selection and order
allocation process. For orange juice product in Quarter three, two sup­
pliers (1 and 4) are selected, when the LSTM network is used. However,
for MLP and SARIMA, three suppliers (1, 4, 5) are chosen. The order
numbers are also diverse for the chosen suppliers using the LSTM
network. Therefore, the forecasting models have impacts on the results
of the supplier selection and order allocation. It is also noticed that the
performance metric of the forecasting models.

6. Managerial insights

The findings of the presented study could support manufacturers in


restructuring their conventional supplier selection & order allocation
strategies to keep up with the current business trends. The outcomes of
Fig. 13. The effects on SSOA with and without the first stage. the numerical experiment highlight the managerial and functional

Table 14 Table 16
The effects of variations in β1 and β2 on the selected suppliers (Product = Lemon The effects of the variations in the forecasting models on the selected suppliers
juice, Quarter = 4). (equal weights).
(β1 , β2 ) s=1 s=2 s=3 s=4 s=5 Suppliers The selected suppliers and the orders for Orange juice in
(β1 = 0.6, β2 = 0.4) Quarter 3, 2022 (Q3)
(0.1, 0.9) 600 239.8
(0.3, 0.7) 600 239.8 LSTM MLP SARIMA
(0.4, 0.6) 600 239.8
(0.6, 0.4) 239.8 600 1 3,362.4 3,400 3,400
(0.7, 0.3) 239.8 600 2
(0.9, 0.1) 239.8 600 3
4 3,200 3,200 3,200
5 453 564.6

Table 15
The performance comparison among the LSTM, SARIMA, and MLP models.
Performance metric LSTM SARIMA MLP

Orange Grapefruit Lemon Orange Grapefruit Lemon Orange Grapefruit Lemon

RMSE 325.7 152.3 298.5 1,885.14 1,131.52 1,884.61 991.75 523.45 951.28
MAPE 1.7 2.2 11.5 5.62 46.75 184.40 3.15 12.25 36.33
Average total error 131.98 456.34 419.7

16
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

implications of the proposed framework. A summary of these implica­ the suppliers. The outcomes of Stage 2 are the weights of the suppliers.
tions is as follows: The results from Stages 1 and 2 have been used in Stage 3, where a novel
multi-objective mixed-integer linear programming model has been
• The developed LSTM network can assist managers to determine developed. Five objective functions are included in the developed
future demands with high accuracy. Without the integration of the optimization model. This developed model has been solved using the
demand forecasting part, accurate SSOA planning may not be weighted-sum technique. The results of Stage 3 are a set of efficient
possible under an uncertain environment. solutions. The trade-offs among different objectives have been analyzed
• The SWOT analysis is used in this study as it provides a better un­ using the value path approach and have been discussed. The results of
derstanding of the strengths and weaknesses of organizations. It also the proposed solution framework show that the selected machine
provides insights on probable threats and future business opportu­ learning technique may have direct impact on both the selected sup­
nities to the DM. In addition, SWOT analysis not only provides the pliers and the allocated orders from them.
weights of internal SSOA criteria, but also generates weight for the There are some limitations to this research. First, the proposed
external criteria which enables a pathway of strategic development framework has been evaluated using only the dataset provided by the
for the decision-making authorities. juice company, and the performance of this framework on other datasets
• Supply chain managers can select the suppliers efficiently using the is yet to be explored. Second, in the optimization model, demand is
proposed framework as it considers both qualitative and quantitative uncertain. However, we may have more uncertain parameters in other
criteria. It is important to note that the variability in the weights of cases. Third, this study has explored three forecasting techniques. There
internal and external criteria may change the suppliers as well as the are other advanced deep learning models which have not been consid­
quantities order from them. ered and presented due to the space limitations.
• This study considers multiple conflicting objectives. By using the There are some potential future study avenues related to this
presented value path approach in this study, the DMs can examine research. The framework can be tested using various datasets from
the tradeoffs among different conflicting objectives and make de­ different business settings. The evaluation results for different types of
cisions based on the organization’s goal. In addition, they can create business data can be compared and analyzed in the future. In addition,
proper order quantity strategies from a wide range of sources using instead of considering demand as an uncertain parameter, more uncer­
the proposed integrated framework. tain parameters (e.g., cost) can be considered in the optimization model.
• Finally, this research considers order allocation along the supplier Demand forecasting based on machine learning and deep learning
selection. Hence, the proposed three-stage model is more appro­ models has been combined with SSOA planning in this research. In the
priate for the production and manufacturing industries rather than future, it will be beneficial to have more advanced deep learning models
the service industries. If order allocation is used in the service in­ and compare the results. As future research, other optimization tech­
dustries (e.g., for a customer call center and related calls), this pro­ niques, such as ε-constraint technique and distance method can be used
posed model can also be used. to solve this problem and compare the results.

7. Conclusions CRediT authorship contribution statement

SSOA planning for a future timeline is a complicated task. The un­ Samiul Islam: Conceptualization, Methodology, Software, Valida­
certainty in the future demand may impact the planning negatively. tion, Writing – original draft, Writing – review & editing. Saman Has­
Demand forecasting using an advanced machine learning technique can sanzadeh Amin: Funding acquisition, Conceptualization, Methodology,
provide a higher forecasting accuracy than the traditional methods and Software, Writing – original draft, Writing – review & editing, Super­
can subside the negative effects on the SSOA planning. Selection of vision. Leslie J. Wardley: Funding acquisition, Conceptualization,
suppliers using qualitative criteria is also a complex problem due to the Methodology, Software, Writing – original draft, Writing – review &
vagueness of human thoughts. Effective supplier selection requires a editing, Supervision.
proper evaluation mechanism for both quantitative and qualitative
criteria.
Declaration of Competing Interest
In this study, three stages have been considered for SSOA planning.
In the first stage, an advanced deep learning model with an LSTM
The authors declare that they have no known competing financial
network has been developed to forecast the future demands according to
interests or personal relationships that could have appeared to influence
the real dataset of the Canadian beverage sector. Then, the developed
the work reported in this paper.
model has been compared with other two machine learning models
namely, Seasonal Auto-Regressive Moving Average (SARIMA) and
Data availability
Multi-Layer Perceptrons (MLP). The performances of these three fore­
casting models have been compared using RMSE and MAPE. The results
Data will be made available on request.
show that the developed LSTM network can reduce the forecasting error
by at least 68.5 %, comparing with the other two considered forecasting
Acknowledgments
models. If there is a fluctuation in the data, the forecast of the traditional
LSTM network may not perform well (He et al., 2021; Li et al., 2022). To
The authors would like to thank the editors and reviewers for the
reduce the effects of fluctuations in the prediction, this modified LSTM
excellent comments that improved the quality of the paper significantly.
network proposes the integration of data decomposer and fluctuation
This research has been supported by a grant from the Faculty of Engi­
normalizer which is new in the literature. Besides, unlike the traditional
neering and Architectural Science, Toronto Metropolitan University,
LSTM network with one Encoder-Decoder, three Encoder-Decoder LSTM
and an Insight Development Grant from the Social Sciences and Hu­
network is used in this research to predict three different data elements
manities Research Council (SSHRC) of Canada.
in parallel. The outcomes of Stage 1 are the future demand values for
three different juice products. To assess the suppliers, a novel fuzzy
SWOT model has been presented in Stage 2 that deals with both quali­ References
tative and quantitative factors under uncertainty. Trapezoidal fuzzy
Abdel-Baset, M., Chang, V., Gamal, A., & Smarandache, F. (2019). An integrated
numbers have been used to deal with the vagueness. The SWOT model neutrosophic ANP and VIKOR method for achieving sustainable supplier selection: A
also allows considering both internal and external qualitative criteria for case study in importing field. Computers in Industry, 106, 94–110.

17
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

Ahmadi, S., & Amin, S. H. (2019). An integrated chance-constrained stochastic model for Hendiani, S., Liao, H., Ren, R., & Lev, B. (2020a). A likelihood-based multi-criteria
a mobile phone closed-loop supply chain network with supplier selection. Journal of sustainable supplier selection approach with complex preference information.
cleaner production, 226, 988–1003. Information Sciences, 536, 135–155.
Alavi, B., Tavana, M., & Mina, H. (2021). A Dynamic Decision Support System for Hendiani, S., Mahmoudi, A., & Liao, H. (2020b). A multi-stage multi-criteria hierarchical
Sustainable Supplier Selection in Circular Economy. Sustainable Production and decision-making approach for sustainable supplier selection. Applied Soft Computing,
Consumption, 27, 905–920. 94, Article 106456.
Alegoz, M., & Yapicioglu, H. (2019). Supplier selection and order allocation decisions Hofer, C., Barker, J., & Eroglu, C. (2021). Interorganizational imitation in supply chain
under quantity discount and fast service options. Sustainable Production and relationships: The case of inventory leanness. International Journal of Production
Consumption, 18, 179–189. Economics, 236, Article 108134.
Alikhani, R., Torabi, S. A., & Altay, N. (2019). Strategic supplier selection under Hosseini, S., & Barker, K. (2016). A Bayesian network model for resilience-based supplier
sustainability and risk criteria. International Journal of Production Economics, 208, selection. International Journal of Production Economics, 180, 68–87.
69–82. Hosseini, S., Morshedlou, N., Ivanov, D., Sarder, M. D., Barker, K., & Al Khaled, A.
Alipour, M., Hafezi, R., Rani, P., Hafezi, M., & Mardani, A. (2021). A new Pythagorean (2019). Resilient supplier selection and optimal order allocation under disruption
fuzzy-based decision-making method through entropy measure for fuel cell and risks. International Journal of Production Economics, 213, 124–137.
hydrogen components supplier selection. Energy, 234, Article 121208. Hosseini, S. M., Paydar, M. M., & Triki, C. (2021). Implementing sustainable ecotourism
Amin, S. H., & Zhang, G. (2013). A three-stage model for closed-loop supply chain in Lafour region, Iran: Applying a clustering method based on SWOT analysis.
configuration under uncertainty. International Journal of Production Research, 51(5), Journal of Cleaner Production, 129716.
1405–1425. Islam, S., & Amin, S. H. (2020). Prediction of probable backorder scenarios in the supply
Aurzada, F. (2020). Large deviations for infinite weighted sums of stretched exponential chain using Distributed Random Forest and Gradient Boosting Machine learning
random variables. Journal of Mathematical Analysis and Applications, 485(2), Article techniques. Journal of Big Data, 7(1), 1–22.
123814. Islam, S., Amin, S. H., & Wardley, L. J. (2021). Machine learning and optimization
Bahadori, M., Hosseini, S. M., Teymourzadeh, E., Ravangard, R., Raadabadi, M., & models for supplier selection and order allocation planning. International Journal of
Alimohammadzadeh, K. (2020). A supplier selection model for hospitals using a Production Economics, 242(108315).
combination of artificial neural network and fuzzy VIKOR. International Journal of Jabbarzadeh, A., Fahimnia, B., & Sabouhi, F. (2018). Resilient and sustainable supply
Healthcare Management, 13(4), 286–294. chain design: Sustainability analysis under disruption risks. International Journal of
Bektur, G. (2020). An integrated methodology for the selection of sustainable suppliers Production Research, 56(17), 5945–5968.
and order allocation problem with quantity discounts, lost sales and varying supplier Jia, R., Liu, Y., & Bai, X. (2020). Sustainable supplier selection and order allocation:
availabilities. Sustainable Production and Consumption, 23, 111–127. Distributionally robust goal programming model and tractable approximation.
Büyüközkan, G., Havle, C. A., & Feyzioğlu, O. (2021). An integrated SWOT based fuzzy Computers & Industrial Engineering, 140, Article 106267.
AHP and fuzzy MARCOS methodology for digital transformation strategy analysis in Jauhar, S. K., Amin, S. H., & Zolfagharinia, H. (2021). A proposed method for third-party
airline industry. Journal of Air Transport Management, 97, Article 102142. reverse logistics partner selection and order allocation in the cellphone industry.
Carmona-Benítez, R. B., & Nieto, M. R. (2020). SARIMA damp trend grey forecasting Computers & Industrial Engineering, 162, Article 107719.
model for airline industry. Journal of Air Transport Management, 82, Article 101736. Karunasingha, D. S. K. (2022). Root mean square error or mean absolute error? Use their
Cavalcante, I. M., Frazzon, E. M., Forcellini, F. A., & Ivanov, D. (2019). A supervised ratio as well. Information Sciences, 585, 609–629.
machine learning approach to data-driven simulation of resilient supplier selection Kaur, H., & Singh, S. P. (2021). Multi-stage hybrid model for supplier selection and order
in digital manufacturing. International Journal of Information Management, 49, 86–97. allocation considering disruption risks and disruptive technologies. International
Chagas, J. B., & Wagner, M. (2022). A weighted-sum method for solving the bi-objective Journal of Production Economics, 231, Article 107830.
traveling thief problem. Computers & Operations Research, 138, Article 105560. Kellner, F., & Utz, S. (2019). Sustainability in supplier selection and order allocation:
Chai, J., & Ngai, E. W. (2020). Decision-making techniques in supplier selection: Recent Combining integer variables with Markowitz portfolio theory. Journal of cleaner
accomplishments and what lies ahead. Expert Systems with Applications, 140, Article production, 214, 462–474.
112903. Kilic, H. S., & Yalcin, A. S. (2020). Modified two-phase fuzzy goal programming
Chaturvedi, S., Rajasekar, E., Natarajan, S., & McCullen, N. (2022). A comparative integrated with IF-TOPSIS for green supplier selection. Applied Soft Computing, 93,
assessment of SARIMA, LSTM RNN and Fb Prophet models to forecast total and peak Article 106371.
monthly energy demand for India. Energy Policy, 168, Article 113097. Kolassa, S. (2020). Why the “best” point forecast depends on the error or accuracy
Chen, Z., Ming, X., Zhou, T., & Chang, Y. (2020). Sustainable supplier selection for smart measure. International Journal of Forecasting, 36(1), 208–211.
supply chain considering internal and external uncertainty: An integrated rough- Li, J., Huang, Y., Li, Q., & Li, Y. (2022). Closed-LSTM neural network based reference
fuzzy approach. Applied Soft Computing, 87, Article 106004. modification for trajectory tracking of piezoelectric actuator. Neurocomputing, 467,
Chen, P., & Sung, S. H. (2022). A Marcinkiewicz-Zygmund type strong law for weighted 379–391.
sums of φ-mixing random variables and its applications. Journal of Mathematical Liou, J. J., Chang, M. H., Lo, H. W., & Hsu, M. H. (2021). Application of an MCDM model
Analysis and Applications, 505(1), Article 125572. with data mining techniques for green supplier evaluation and selection. Applied Soft
Du, B., Guo, S., Huang, X., Li, Y., & Guo, J. (2015). A Pareto supplier selection algorithm Computing, 107534.
for minimum the life cycle cost of complex product system. Expert Systems with Liu, A., Xiao, Y., Lu, H., Tsai, S. B., & Song, W. (2019b). A fuzzy three-stage multi-
Applications, 42(9), 4253–4264. attribute decision-making approach based on customer needs for sustainable
Dubey, A. K., Kumar, A., García-Díaz, V., Sharma, A. K., & Kanhaiya, K. (2021). Study supplier selection. Journal of Cleaner Production, 239, Article 118043.
and analysis of SARIMA and LSTM in forecasting time series data. Sustainable Energy Liu, J. G., Yang, X. J., Feng, Y. Y., & Cui, P. (2020). New fractional derivative with
Technologies and Assessments, 47, Article 101474. sigmoid function as the kernel and its models. Chinese Journal of Physics, 68,
Esmaeili-Najafabadi, E., Nezhad, M. S. F., Pourmohammadi, H., Honarvar, M., & 533–541.
Vahdatzad, M. A. (2019). A joint supplier selection and order allocation model with Liu, L., Liu, J., Zhou, Q., & Huang, D. (2022). Machine learning algorithm selection for
disruption risks in centralized supply chain. Computers & Industrial Engineering, 127, windage alteration fault diagnosis of mine ventilation system. Advanced Engineering
734–748. Informatics, 53, Article 101666.
Esmaeili-Najafabadi, E., Azad, N., & Nezhad, M. S. F. (2021). Risk-averse supplier Liu, Y., Eckert, C., Yannou-Le Bris, G., & Petit, G. (2019a). A fuzzy decision tool to
selection and order allocation in the centralized supply chains under disruption risks. evaluate the sustainable performance of suppliers in an agrifood value chain.
Expert Systems with Applications, 175, Article 114691. Computers & Industrial Engineering, 127, 196–212.
Feng, J., & Gong, Z. (2020). Integrated linguistic entropy weight method and multi- Lu, Z., Sun, X., Wang, Y., & Xu, C. (2019). Green supplier selection in straw biomass
objective programming model for supplier selection and order allocation in a industry based on cloud model and possibility degree. Journal of cleaner production,
circular economy: A case study. Journal of Cleaner Production, 277, Article 122597. 209, 995–1005.
Fernandes, I. F., Goldbarg, E. F., Maia, S. M., & Goldbarg, M. C. (2021). Multi-and many- Lyu, M. Z., Wang, J. M., & Chen, J. B. (2021). Closed-form solutions for the probability
objective path-relinking: A taxonomy and decomposition approach. Computers & distribution of time-variant maximal value processes for some classes of Markov
Operations Research, 133, Article 105370. processes. Communications in Nonlinear Science and Numerical Simulation, 99, Article
Firouzi, F., & Jadidi, O. (2021). Multi-objective model for supplier selection and order 105803.
allocation problem with fuzzy parameters. Expert Systems with Applications, 180, Medhi, P. K., & Mondal, S. (2016). A neural feature extraction model for classification of
Article 115129. firms and prediction of outsourcing success: Advantage of using relational sources of
Fonseca, J. D., Commenge, J. M., Camargo, M., Falk, L., & Gil, I. D. (2021). Sustainability information for new suppliers. International Journal of Production Research, 54(20),
analysis for the design of distributed energy systems: A multi-objective optimization 6071–6081.
approach. Applied Energy, 290, Article 116746. Memari, A., Dargi, A., Jokar, M. R. A., Ahmad, R., & Rahim, A. R. A. (2019). Sustainable
Gomez-Biscarri, J., & Hualde, J. (2015). A residual-based ADF test for stationary supplier selection: A multi-criteria intuitionistic fuzzy TOPSIS method. Journal of
cointegration in I (2) settings. Journal of Econometrics, 184(2), 280–294. manufacturing systems, 50, 9–24.
Hasan, M. M., Jiang, D., Ullah, A. S., & Noor-E-Alam, M. (2020). Resilient supplier Mina, H., Kannan, D., Gholami-Zanjani, S. M., & Biuki, M. (2021). Transition towards
selection in logistics 4.0 with heterogeneous information. Expert Systems with circular supplier selection in petrochemical industry: A hybrid approach to achieve
Applications, 139, Article 112799. sustainable development goals. Journal of Cleaner Production, 286, Article 125273.
Hasani, A., Mokhtari, H., & Fattahi, M. (2021). A multi-objective optimization approach Miranda-Ackerman, M. A., Azzaro-Pantel, C., & Aguilar-Lasserre, A. A. (2017). A green
for green and resilient supply chain network design: A real-life Case Study. Journal of supply chain network design framework for the processed food industry: Application
Cleaner Production, 278, Article 123199. to the orange juice agrofood cluster. Computers & Industrial Engineering, 109,
He, K., Ji, L., Wu, C. W. D., & Tso, K. F. G. (2021). Using SARIMA–CNN–LSTM approach 369–389.
to forecast daily tourism demand. Journal of Hospitality and Tourism Management, 49,
25–33.

18
S. Islam et al. Computers & Industrial Engineering 174 (2022) 108825

Mishra, A. R., Rani, P., & Prajapati, R. S. (2021). Multi-criteria weighted aggregated sum Schramm, V. B., Cabral, L. P. B., & Schramm, F. (2020). Approaches for supporting
product assessment method for sustainable biomass crop selection problem using sustainable supplier selection-A literature review. Journal of cleaner production,
single-valued neutrosophic sets. Applied Soft Computing, 113, Article 108038. 123089.
Mohammed, A., Harris, I., & Govindan, K. (2019). A hybrid MCDM-FMOO approach for Singh, D., & Singh, B. (2020). Investigating the impact of data normalization on
sustainable supplier selection and order allocation. International Journal of Production classification performance. Applied Soft Computing, 97, Article 105524.
Economics, 217, 171–184. Statista Juice Report. (2021). The market for Non-Alcoholic Drinks is structured into
Moheb-Alizadeh, H., & Handfield, R. (2018). An integrated chance-constrained retail sales for at home consumption and on-premises or foodservice sales for out-of-
stochastic model for efficient and sustainable supplier selection and order allocation. home consumption. Retrieved from https://www.statista.com/study/48822/juices-
International Journal of Production Research, 56(21), 6890–6916. report/. Accessed January 27, 2022.
Moheb-Alizadeh, H., & Handfield, R. (2019). Sustainable supplier selection and order Statistics Canada. (2021a). Table 18–10-0266-02. Industrial product price index, by
allocation: A novel multi-objective programming model with a hybrid solution product, percentage change, monthly., 2022.
approach. Computers & industrial engineering, 129, 192–209. Statistics Canada. (2021b). Table 32-10-0053-01 Supply and disposition of food in
Mohebalizadehgashti, F., Zolfagharinia, H., & Amin, S. H. (2020). Designing a green Canada (x 1,000). Retrieved from https://www150.statcan.gc.ca/t1/tbl1/en/tv.
meat supply chain network: A multi-objective approach. International Journal of action?pid=3210005301. Accessed January 27, 2022.
Production Economics, 219, 312–327. Tavana, M., Fallahpour, A., Di Caprio, D., & Santos-Arteaga, F. J. (2016). A hybrid
Mojumder, A., & Singh, A. (2021). An exploratory study of the adaptation of green intelligent fuzzy predictive model with simulation for supplier evaluation and
supply chain management in construction industry: The case of Indian Construction selection. Expert Systems with Applications, 61, 129–144.
Companies. Journal of Cleaner Production, 295, Article 126400. Tirkolaee, E. B., Mardani, A., Dashtian, Z., Soltani, M., & Weber, G. W. (2020). A novel
Naqvi, M. A., & Amin, S. H. (2021). Supplier selection and order allocation: a literature hybrid method using fuzzy decision making and multi-objective programming for
review. Journal of Data, Information and Management, 1-15. sustainable-reliable supplier selection in two-echelon supply chain design. Journal of
Nepal, B., & Yadav, O. P. (2015). Bayesian belief network-based framework for sourcing Cleaner Production, 250, Article 119517.
risk analysis during supplier selection. International Journal of Production Research, Ventura, J. A., Bunn, K. A., Venegas, B. B., & Duan, L. (2021). A coordination mechanism
53(20), 6114–6135. for supplier selection and order quantity allocation with price-sensitive demand and
Nasr, A. K., Tavana, M., Alavi, B., & Mina, H. (2021). A novel fuzzy multi-objective finite production rates. International Journal of Production Economics, 233, Article
circular supplier selection and order allocation model for sustainable closed-loop 108007.
supply chains. Journal of Cleaner Production, 287, Article 124994. Wang, P., Hafshejani, B. A., & Wang, D. (2021). An improved multilayer perceptron
OFVGA Directory. (2021). Ontario fruits and vegetables growers’ association. Retrieved approach for detecting sugarcane yield production in IoT based smart agriculture.
from https://www.ofvga.org/overview. Accessed January 27, 2022. Microprocessors and Microsystems, 82, Article 103822.
Orji, I. J., & Ojadi, F. (2021). Investigating the COVID-19 pandemic’s impact on Wang, R., Li, X., & Li, C. (2021). Optimal selection of sustainable battery supplier for
sustainable supplier selection in the Nigerian Manufacturing Sector. Computers & battery swapping station based on Triangular fuzzy entropy-MULTIMOORA method.
Industrial Engineering, 107588. Journal of Energy Storage, 34, Article 102013.
Ouadfel, S., & Abd Elaziz, M. (2021). A multi-objective gradient optimizer approach- Wen, T. C., Chang, K. H., & Lai, H. H. (2020). Integrating the 2-tuple linguistic
based weighted multi-view clustering. Engineering Applications of Artificial representation and soft set to solve supplier selection problems with incomplete
Intelligence, 106, Article 104480. information. Engineering Applications of Artificial Intelligence, 87, Article 103248.
Papen, P., & Amin, S. H. (2019). Network configuration of a bottled water closed-loop Wong, J. T. (2020). Dynamic procurement risk management with supplier portfolio
supply chain with green supplier selection. Journal of Remanufacturing, 9(2), selection and order allocation under green market segmentation. Journal of Cleaner
109–127. Production, 253, Article 119835.
Qaiser, I. (2022). A comparison of renewable and sustainable energy sector of the South Wu, C., Lin, Y., & Barnes, D. (2021). An integrated decision-making approach for
Asian countries: An application of SWOT methodology. Renewable Energy, 181, sustainable supplier selection in the chemical industry. Expert Systems with
417–425. Applications, 184, Article 115553.
Rashidi, K., Noorizadeh, A., Kannan, D., & Cullinane, K. (2020). Applying the triple Yin, J., Ning, C., & Tang, T. (2022). Data-driven models for train control dynamics in
bottom line in sustainable supplier selection: A meta-review of the state-of-the-art. high-speed railways: LAG-LSTM for train trajectory prediction. Information Sciences,
Journal of Cleaner Production, 269, Article 122001. 600, 377–400.
Sanito, R. C., You, S. J., Chang, T. J., & Wang, Y. F. (2020). Economic and environmental Yousefi, S., Jahangoshai Rezaee, M., & Solimanpur, M. (2021). Supplier selection and
evaluation of flux agents in the vitrification of resin waste: A SWOT analysis. Journal order allocation using two-stage hybrid supply chain model and game-based order
of Environmental Management, 270, Article 110910. price. Operational Research, 21(1), 553–588.
Sarkis, J., & Dhavale, D. G. (2015). Supplier selection for sustainable operations: A triple- Zandieh, M., & Aslani, B. (2019). A hybrid MCDM approach for order distribution in a
bottom-line approach using a Bayesian framework. International Journal of Production multiple-supplier supply chain: A case study. Journal of Industrial Information
Economics, 166, 177–191. Integration, 16, Article 100104.
Sazvar, Z., Tafakkori, K., Oladzad, N., & Nayeri, S. (2021). A capacity planning approach Zhang, L., Zhang, Q., Wu, J., Liu, Y., Yu, L., & Chen, Y. (2022). Moisture detection of
for sustainable-resilient supply chain network design under uncertainty: A case study single corn seed based on hyperspectral imaging and deep learning. Infrared Physics
of vaccine supply chain. Computers & Industrial Engineering, 159, Article 107406. & Technology, 104279.
Schmidt-Hieber, J. (2020). Nonparametric regression using deep neural networks with
ReLU activation function. The Annals of Statistics, 48(4), 1875–1897.

19

You might also like